Player is loading...

Embed

Embed code

Transcriptions

Note: this content has been automatically generated.
00:00:00
i think yeah much uh e. one else for for give me the chance to present here
00:00:04
and again all what work is um person on trial is
00:00:09
um so we billing and learning algorithms which we
00:00:12
use uh apply x. rays to extract certain information what
00:00:16
prediction about what all the classification of the disease
00:00:21
um i think i have to give a lot utilities numbers uh
00:00:25
and summer you're is it's it's a lot of patients being effect but is uh try to
00:00:30
a to a tight is you know it has general is r. a. or all way
00:00:34
and he likes and on what these diseases all have in common and is that that imaging
00:00:40
uh is performed usually easy to hand is the knee is it um any other extremity um
00:00:46
and what we have seen here um my my teammate image types of that is that
00:00:51
um this whole is x. ray diagnoses uh which are
00:00:54
performed by physicians are usually quite subjective underperformed manually
00:00:59
i'm very often depending on the experience of decision was off that
00:01:02
day you know and time that that uh analysis being performed
00:01:07
and it's very often results at least what we've seen the case of last friday's in a
00:01:12
quite nonspecific treatment and we believe that um they are always out there to change that
00:01:18
so basically if you look at our applications let's say the field off you know lot testing than automation
00:01:25
automated analyses is a standard right you wouldn't find anyone looking you know that incoming cells anymore
00:01:32
um but unfortunately if you look at a few of radiology not peaks
00:01:37
and the way how be assessed x. rays nowadays is um as one of our
00:01:42
our doctors uses said i i was the said it's it's almost like a hundred years ago
00:01:46
so with the x. ray and we come up with a conclusion and medication comes back
00:01:50
you two years later began more less you look at it and then try to find out what has changed
00:01:55
and our approach was um to um kind of like a work from um to change it actually is that
00:02:01
we are actually can look into from a really active model which means that the patient comes to us
00:02:06
uh we look at the x. ray and we say well there is something i come back
00:02:09
actually when you know it's getting worse um that we go more to a proactive model
00:02:15
well we actually say well they are certain signs here which put it as a risk factor
00:02:20
are there certain anomalies in the x. ray basin that we started preventive treatments much sooner
00:02:26
uh or exact final exam would be retractable question much more
00:02:29
accurately did we say well you wait what's that direction
00:02:33
uh and based on that we can actually do something about that right and just waiting until um you know the page when it comes
00:02:39
to us with we see a pain so basically does what we are
00:02:44
doing um he again on the sample often e. x. ray
00:02:48
um where we actually use 'em weighting schemes like a lawrence score
00:02:52
for example and again this is your focusing on osteoarthritis
00:02:56
but we take the x. ray instead of doing it manually design and um we have a
00:03:00
software trained a base on the learning machine learning as we heard the previous talks
00:03:07
actually i take 'em all these men of measurements of the catalan score and put
00:03:12
it into a detail disease report um within you just if you sex actually
00:03:18
so um is you just screen job as screen shot of a recent the integration of our software
00:03:24
so basically you see here um the pain protocol where you see on the right side uh we just study
00:03:30
uh and then you see here a feature the the nice then hear the annotations by the software
00:03:35
um basically um we we remove that information initially they were like all kinds
00:03:40
of different from a she about you know ah suffice colour rose
00:03:43
in the likes so we actually kept quite simple based on the phoebe
00:03:47
from the radiologists which just show um the cable on score
00:03:50
and then there's the report here with all the the the uh was
00:03:54
equating the likes which i will go to multichannel next slide
00:03:59
so what is the output of um that software again you see the final played catalan
00:04:05
score for each of you know if he's a bilateral needs for each side
00:04:09
uh yeah he was a great as well and so it's a combination of cattle lawrence and was it
00:04:15
um then you have here in the lower part the report a great detail
00:04:19
measurements and and what we measure here are different and anatomical parameters
00:04:24
a joint space with joint space area which is quite indirect measurement of the cottage area
00:04:30
uh and what we ought to do here uh we have twenty algorithm to ah chi like i'm i'm
00:04:35
kinda calibrated for the distance you know the difference is a distance right so if a patient
00:04:41
stay still closed the tact or for out all the does have an impact on the measurements
00:04:45
and what we trained our software on was to compensate for these
00:04:49
measurements have uh what we call the standardised quantitative measurements
00:04:54
unless the of course uh what our goal was to provide to position a
00:04:59
a comparison to uh this case it would be left to right knee
00:05:02
but of course going for guide you would be that um the we all
00:05:06
just get the x. ray this can swing form and what your
00:05:09
feeling would get the end would be um that's and i support here and if the patient comes back a year two years later
00:05:15
um you can compare again you know either the final kale kale school or or uh
00:05:21
the sops course uh for joint space now rings the roses and austin fights
00:05:27
so here are just a some performance i cover sticks uh this offer is currently undergoing f.
00:05:32
d. a. approval so we went for all these uh validation sets uh and the likes
00:05:37
uh so for the confusion matrix is here the cattle lawrence great of course to detect if a
00:05:42
patient has away another way is um is um you know quite good the accuracy sensitivity specificity
00:05:50
where it becomes tricky as we heard a previously is especially in the early um
00:05:54
uh early some uh uh some of the disease scale one and two
00:05:59
we can see here um so that you know the two living here you see it it's
00:06:04
it's quite difficult to differentiate between one and two uh here especially scary here
00:06:08
but then when we go out for up to kayla three and four
00:06:12
uh the softer becomes much better in the green with with the ground truth actually
00:06:16
and this is based on the data set of roughly four thousand patients uh for your follow ups
00:06:23
so roughly fifteen thousand images which we collected a large uh us uh um multi centre study
00:06:31
we can perform there we decided so we said okay you know the software's here um we know that there is
00:06:37
a a you didn't rape can position so does a soft actually change something in terms of the grid positions
00:06:44
uh for this purpose we uh selected a random sample of her thirty six x. rays
00:06:49
we gave to position the plane x. ray and then we gave to position the same x. ray of course in different order
00:06:55
with the softer generate a report next it and ask position you do you agree with that or not
00:07:02
uh and basically see here um based on on the couple score is that you know
00:07:08
the the undated agreement between that with the specialists is here mark in in blue
00:07:14
um and then when they had the software with it then it was you know we get us um and improvement
00:07:21
in in each of these categories with exactly germans was nearing with a already agree without the softer very good
00:07:27
so was as a gift improvement each of these categories when a position actually
00:07:30
had to hang protocol at realised um report next to the x. ray
00:07:36
then uh with a big women rate um how do you agree uh between each
00:07:40
other and you see here the the blue circle one is the unaided uh
00:07:44
version and then if we gave them the the the report annex
00:07:49
it then uh we also saw that and still improvement
00:07:52
in a pretty much each of the categories so in other words and to position had something extreme where software
00:07:58
kinda gave you the every indication then uh and they start to agree more uh with each other
00:08:05
and initially the initial trial we had with k. loans wars that um the only get thirty percent
00:08:10
of cases so um so that was actually reading the we just study a big improvement
00:08:15
as to this is of course weren't you know fully blinded to what they are actually should uh should be doing
00:08:22
i'm not this is really just um to support uh without is of predicts the way you know you wanna
00:08:28
great certain common feature of the x. ray but not a question is can we also try this information
00:08:34
can we keep the decision information where they say well it might became learns to um
00:08:39
but you know are you the high risk of developing or you're not away
00:08:43
um and for that purpose our initial focus was on text analysis to use that as a
00:08:49
as an additional uh the data point either source in the uh in the image analysis
00:08:54
and actually texan else's is something which do not differ quite sometime most being used in the field of
00:08:59
false crosses detect uh the risk of a factor so we did basically we looked at um
00:09:06
again at a large uh us study where we had a baseline and follow up a data available
00:09:12
that and want to know um if we have a a base like with patients um
00:09:17
what what is the accuracy we can uh we can achieve to
00:09:21
then if i within these baseline groups all kills zero um
00:09:25
that you know how many of these patients actually will have kale large in one um in eighty four month follow up
00:09:32
today was the i'd hear a um and what we did then we build different models
00:09:37
look at uh things like a clinical risk factors section houses joint
00:09:41
space with with a lot and so on so on
00:09:44
and imply that all the results um and um what we see here is we
00:09:48
say five of male and female we saw difference between these two groups
00:09:52
um what you see here the blue line actually is here on on the look it females is the the
00:09:58
standard model way just have that little factors joint space with p. m. i. h. and and the likes
00:10:03
then if you just look at text analysis um so if you look at now um the
00:10:08
bone mike architecture as a as a i would say then the production mode likes
00:10:13
increased in other words um we went up from legacy of predicting which patient will be
00:10:18
in the progress or group from zero point six seven two zero point seven eight
00:10:22
and if we combine text analysis uh see here uh at the bottom
00:10:27
i take announces with him i joined space with and all the other factors
00:10:31
we achieved next c. of your point eight in predicting which of these patients in the
00:10:36
k. l. zero group actually we'll have yellow one higher identity foreman for awhile
00:10:42
the male group uh perform worse and we believe the main reason ever is that we
00:10:47
just had you know a very small sample size uh from from male patients
00:10:52
the main issue with protection else's though is that it's quite sensitive to the image parameters um
00:10:59
so you know playing on the large data set with different eczema dell is kinda tricky
00:11:04
so what we did as far as the next that actually was look at alternative such as the planning for example
00:11:10
uh and we use a day the same data set and uh
00:11:13
we were able to achieve almost the same results already
00:11:17
um so we're at seventy eight percent accuracy so far um and if you just compare that to
00:11:22
uh the work so i think a static little later so be my joint space age
00:11:28
um we already are we are performing the standard model with our keep learning model
00:11:32
which has the benefit that it actually tells you also which of the regions you're in the x. say what you see here is blue
00:11:39
patches you actually the areas in the x. ray pose the each individual patients
00:11:44
what a net but actually take it that he's either the yeah areas we're
00:11:49
at a form of follow up they will be a change happened
00:11:53
and so it it it's quite uh quite um and need to result here we
00:11:56
gonna presented at u. c. r. and in two weeks actually the or presentation
00:12:02
um but what other tools have we have to look that way half we can
00:12:06
apply these these these these these networks these people networks on other applications
00:12:11
i'm and r. a. of course we had already met previous presentation is something which very interesting
00:12:17
and again the same approach as with oscar tried is uh the first there was to use text analysis
00:12:23
to see if we can use text analysis to extract certain information or the
00:12:26
x. ray which helps their remote apologies to predict or the the progression
00:12:31
uh and initial results here against small a small sample size where a promising um
00:12:38
so so we are now expanding on that data set we have a another these which is much larger
00:12:44
and we do basically here is um in the first step is to automate the input data
00:12:49
um so basically um we have train the net work to um you know measure joint space with for
00:12:55
example and also detect contours which we don't data uh use for actually um the rose detection
00:13:02
uh and this this is just a mock up this this report you don't be some pay attention
00:13:06
it's really just what what the software does what we have tried the software on to detect
00:13:10
um then again um but we can only do we look at the sharp
00:13:14
score um to see um can we actually train a network to detect
00:13:19
on on which of the joints actually there is our already progression happening
00:13:24
uh and and of course uh the next it would be to use that to predict are a but that's where we are
00:13:29
right now so far it's really just a message on space with uh and to really assess the shark score automatically
00:13:37
um and application um is really just it's all it's all the same
00:13:40
concept basically to use it people networks and just give them instead
00:13:44
of hip a hand or a instead of any hand or hip
00:13:47
this case for example uh we look at catalogues measurements indicate
00:13:52
uh the first again is measured joint space with and then go again
00:13:56
uh also fights uh as closes all those kind of things
00:14:00
um and then you know what is and uh it was a nice ipods was
00:14:03
all the match automatic angles so you know the softer actually detects you
00:14:07
also contour second place an ad for the angles here so you know do
00:14:11
it all just doesn't have to mentally i click on the x. ray
00:14:16
um and then the last pet project it doesn't really fit so much it it is a it is topic but it's
00:14:20
really something which we work on just a month ago so work on that we're look at bone age assessment
00:14:25
um and um we have done we spend so much work in a few of
00:14:29
austria fridays um that we just thought you know hey now we have hand
00:14:33
now if your hands can be actually used what we have falsified is all the different with different feel
00:14:38
and just within a month actually i we were able to train at work to uh assess people in h. a.
00:14:45
uh in um you know with a with a with a with a range of plus minus five months
00:14:50
next this was um the winning cattle competition at our slate was eighteen actually was that range
00:14:56
so that shows you that showed us that you know if we have the proper image data
00:15:00
or you can apply these algorithms to almost any you know any any any data or
00:15:05
um but what is all important and what i must stress again we were before is it's all the input data
00:15:10
and this was it they're nice sample set with very clean input data are and
00:15:14
if you don't have that um you will not was a achieve such results
00:15:20
yeah in conclusion and the outlook reform from our point of
00:15:23
view is that these these people earning oh digital um
00:15:27
the technology is uh actually can provide crime savings by
00:15:31
automate automating uh so far pretty mellow process
00:15:35
we don't see the change of work was necessary these things one the background now we really became mature he you
00:15:41
know half ago two years ago it was very tedious now
00:15:44
these things um our um have become like grownups
00:15:48
um this whole solicitation an object reporting really we have seen does actually have
00:15:53
an impact so we think that is really adds value to it
00:15:56
um outcome measures is something which would you quite often do we know what
00:16:00
you mean is working not how can we actually track um the effectiveness
00:16:05
um things like a second opinion i'm in breast cancer these things are very common
00:16:10
innovative of pubic somewhat taller gee we're not there yet but you believe that you know
00:16:14
the time is ripe with these things also um our news more often that field
00:16:20
keep track of treatment changes i mentioned for sure and unless of course increase agreement between
00:16:26
readers and when i saw for your provides a a first reading on this
00:16:31
yeah with this i wanna kind of presentation and um
00:16:34
i think you can for time thank you how
00:16:39
honking which are a a question from the audience
00:16:46
a half want to start a new one just that he was saying comparing different model
00:16:52
it was like a each joint be emma it's on and then just
00:16:56
like uh was texture the so that was the first plaza second
00:17:01
um i'm i was quite surprised to see that combining the first and second
00:17:06
doesn't really improve significantly easy outcome because you move a a you see from seventy
00:17:11
eight to eighty which is clearly not significant uh you don't have the
00:17:16
confidence interval that pretty sure yep so so does it mean that most of
00:17:21
the performance agree that budget texture and the rest is just like
00:17:27
minor adults or does it mean that a sample size or how would you
00:17:32
explain that because maybe the focus should be on the texture only
00:17:36
and the rest is just okay now i have but not ready of added that yeah no i i do that with you
00:17:42
and and that was why we showed it was because we can like no question if faction as we add that much value
00:17:49
but we've seen that i'm actually really i provide information which you know the human eye um conflict or the x. ray
00:17:56
so we think that this certainly is the way to go and the reason i moved all the loading
00:18:01
is because we find it difficult to calibrate you know that images according to set the standard
00:18:07
and we've seen that with deep learning on the algorithms are the stage where
00:18:12
could easier work with x. wasn't even without this but if they would be a single standard in x.
00:18:18
rays then text analysis that we would be superior to most of the current her measurements we're using
00:18:27
so thank you very much is a pretty brilliant presentation very nice and uh i
00:18:31
have a particular interest in house or try to so actually want to ask
00:18:34
you is so from the most study what if an animal studies they actually
00:18:38
also look at at bomber lesions and mostly i'm right imaging obviously we
00:18:42
i song or high but not not most sorry about the most l. c. look
00:18:46
at us undivided okay yeah sorry yeah yeah okay yeah so i was
00:18:50
wondering so they are they found actually that were some and arrive features that
00:18:54
that they found predictive for development of radio graphical way later on
00:18:58
i was wondering how does this compare to your bone textual measurements can
00:19:02
you detected pretty early on and a second question would be
00:19:05
when using a production algorithm or you saw that most of the changeable taking place that either at the middle of the t. v. or
00:19:12
or at the edge of the of the bone but not in the central part where
00:19:15
normally would have a a college last do you have any uh explanation for that
00:19:20
um so so the first question um so we we just we're just signs look now into into these correlations
00:19:27
um what the work with the right now is to um actually look into
00:19:31
can we um predict or detectable male lesions on conventional x. rays
00:19:37
so in the away i study for example they had that correlations also you know we're trying
00:19:41
to detect a certain area based on an x. rating keep that back to that position
00:19:46
um i think is a very interesting um research question which we're looking wrote
00:19:50
to right now i hope i have no more results sometimes you know
00:19:54
we end this year um to the uh to the uh second question
00:19:59
i can give you an explanation uh why's that uh we you know if that old information to network and then
00:20:05
it's the little black box right they give you something out and you don't really know why these like that
00:20:10
so uh we need understand more you know the better way why that came that conclusion
00:20:15
and i think the key that will be the chip we have um more image data sets like you i like the most to repeat that and see
00:20:21
if these results are actually know almost same different patient populations for example
00:20:30
could be could be yes i've just the previous question um with the with the shop score this
00:20:36
shining pictures with the what is this a heat that yes and i'm just i don't know
00:20:42
what you think but for me when we see something shining on an x. ray or c. t.
00:20:47
we think about aspects eighteen so i'll still plus or packed a a close up take
00:20:54
but but this is beta thing right this shining it's it's
00:20:57
confusing me but amazingly it's and it makes me what
00:21:02
does this do with us how it it's that it's i'm
00:21:06
coming from and more like data science background um
00:21:09
those are concert would you use for different applications so on
00:21:13
but we still need to do to understand what the decision in what kind of formed a physician would like that information
00:21:19
the deal was really just to give him his position id or um on
00:21:23
the house you've your discharge was based on uh he met right um
00:21:28
i'm not too happy about that because we on it was it it's an specific matter right
00:21:33
um so i think they need to be more input from physicians well product exact i think that's a clinician i
00:21:38
would like to have this in future he met matt on the x. ray to see where is the emotion
00:21:45
sit here on the still a process or maybe i am check there for example
00:21:50
so uh that's something we need to yeah okay so so i can yeah i that
00:21:53
that's that's what we're not actually sure about with that with them that way production
00:21:58
and then and that you will works there we use the same concept with heat network show these
00:22:03
are the regions where it will compress the future above are a we actually just early on
00:22:10
alright so thank you very much so ah ha ha
00:22:17
that's why we're here to uh what uh what might go
00:22:21
from copenhagen i don't read the name of university because
00:22:26
um is too complicated for me i'm sorry i don't want to uh to be rude on that plane
00:22:31
and uh she would present a electro so he has one letter g.

Share this talk: 


Conference program

Demystification the digital health world
Thomas Hügle
1 Feb. 2019 · 9:41 a.m.
200 views
Digital lessons learned from musculoskeletal radiology
Patrick Omoumi
1 Feb. 2019 · 11:37 a.m.